Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.568276
Title: Software-based approximate computation of signal processing tasks
Author: Anastasia, D.
Awarding Body: University College London (University of London)
Current Institution: University College London (University of London)
Date of Award: 2012
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
This thesis introduces a new dimension in performance scaling of signal processing systems by proposing software frameworks that achieve increased processing throughput when producing approximate results. The first contribution of this work is a new theory for accelerated computation of multimedia processing based on the concept of tight packing (Chapter 2). Usage of this theory accelerates small-dynamic-range linear signal processing tasks (such as convolution and transform decomposition) that map integers to integers, without incurring any accuracy loss. The concept of tight packing is combined with incremental computation that processes inputs in a bitplane-by-bitplane manner (Chapter 3), thereby leading to substantial throughput/distortion scalability within filtering, transform-decomposition and motion-estimation tasks. This framework also provides for region-of-interest computation and has inherent robustness to arbitrary termination of processing, imposed, for example, by a task scheduler. Finally, the concept of packed processing is extended to floating-point (lossy) matrix computations, with particular focus on the generic matrix multiplication (GEMM) routine of BLAS-3 (Chapters 4 and 5). This routine is a fundamental building block for several linear algebra and digital signal processing systems, such as face recognition and neural-network training for metadata-based retrieval systems. In order to compete with the best-performing software designs for GEMM, an implementation using single instruction, multiple data (SIMD) instructions is presented and analyzed. The proposed approach demonstrates substantial performance scaling in practice; specifically, it is shown to achieve up to twice the processing throughput of the best designs for GEMM when producing approximate results (under the same hardware). In summary, the proposed approximate computation of signal processing tasks can be selectively disabled thereby producing conventional full-precision/lower-throughput processing when deemed necessary. Importantly, the proposed software designs run on off-the-shelf computer hardware and provide for on-demand reconfiguration, depending on the input data and the precision specification (from full precision to noisy computation). Thus, the proposed approximate computation framework allows for backward compatibility and can be offered as an add-on service, creating significant competitive advantages for application developers. It can be used in mobile or high-performance computing systems when the precision of computation is not of critical importance (error-tolerant systems), or when the input data is intrinsically noisy.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.568276  DOI: Not available
Share: